A GPT-4 Reticular Chemist for Guiding MOF Discovery

Results and Discussion

• Four isoreticular MOF-521 compounds were discovered through the GPT-4 Reticular Chemist workflow, with synthesis, linker preparation, optimization, and characterization designed by GPT-4 and executed by a human researcher.MOF-521-H, -oF, -mF, and -CH3 share the generic formula [Al3(μ-OH)3(HCOO)3(BTB-X)].Source: Results and Discussion · pp. 3–4
• The three-phase workflow uses ChemScope for project blueprints, ChemNavigator for task recommendations based on progress, and ChemExecutor for detailed experimental steps.ChemNavigator incorporates the human feedback loop; the workflow is developed and operated in natural language without coding.Source: Figure 1 · p. 2, Results and Discussion · p. 3

Introduction

• Human researchers report successes and failures after experiments, allowing GPT-4 to update its in-context memory and adapt subsequent recommendations.Researchers choose among suggested actions based on their merit, available resources, instruments, and expertise.Source: Results and Discussion · p. 3, Introduction · p. 3

Results and Discussion

• The MOF-521 compounds share a rod-SBU structure and a topology represented as a 2-nodal net; the authors report this as the first MOF with that topology.The three rod SBUs are closely packed, with a 6.54 Å distance between their central axes, and formates block coordination sites between them.Source: Results and Discussion · p. 6
• Nitrogen sorption confirmed permanent porosity across the MOF-521 family, with BET surface areas of 1696 m2 g–1 for MOF-521-H, 1535 m2 g–1 for MOF-521-oF, 1562 m2 g–1 for MOF-521-mF, and 1311 m2 g–1 for MOF-521-CH3.The authors report Type I isotherms and suggest central-benzene substitution can influence and potentially tune porosity.Source: Results and Discussion · p. 6, Table 2 · p. 7
• Prompt iteration counts were similar across the four compounds despite differences in their discovery routes, while the later three required fewer iterations than the initial MOF-521-H.The authors attribute the later reduction to MOF-521-H's compiled successes and failures serving as guidance for subsequent compounds; over 80% of GPT-4 task recommendations aligned with researchers' immediate challenges.Source: Results and Discussion · pp. 7–8

Conclusion

• The authors present the framework as a systematic human-AI collaboration and a potential basis for broader scientific applications, while describing GPT-4 as supporting rather than replacing human judgment.They identify full automation and robotization of experimental steps as a future direction.Source: Conclusion · p. 8, Results and Discussion · p. 8

Supporting Information

• The author list names Zhiling Zheng, Zichao Rong, Nakul Rampal, Christian Borgs, Jennifer T. Chayes, and Omar M. Yaghi.Omar M. Yaghi is marked as the corresponding author.Source: Supporting Information · p. 10
• The affiliations include UC Berkeley’s Department of Chemistry and Kavli Energy Nanoscience Institute.The Bakar Institute of Digital Materials for the Planet is also listed.Source: Supporting Information · p. 10
• Additional listed Berkeley affiliations span electrical engineering and computer sciences, mathematics, statistics, and the School of Information.The KACST–UC Berkeley Center of Excellence for Nanomaterials for Clean Energy Applications is also listed.Source: Supporting Information · p. 10

Large Language Model

• GPT-4 was used through the ChatGPT web portal, with the March 14, 2023 model version, to avoid requiring coding experience.The study identifies the underlying model as gpt-4-0314.Source: Section S1. Materials and Methods · p. 11

Reticular ChemScope; Reticular ChemNavigator; Reticular ChemExecutor

• ChemScope builds a staged project blueprint, ChemNavigator proposes tasks using project feedback, and ChemExecutor supplies task steps and a feedback template.Each phase is reinitiated with relevant project or task context.Source: Section S1. Materials and Methods · pp. 11–12

Prompt Components

• Prompts combine role and goal instructions, brief memory of prior trials, evaluation, and three structured task choices; memory is shortened because of token limits.The structured choices accommodate practical constraints such as resources and time schedules.Source: Section S1. Materials and Methods · p. 12

Interactive Prompt Refinement

• Prompt drafts were refined through GPT-4 clarification, separate-instance testing, and human feedback until users were satisfied with performance.The full refined prompts are provided in Supporting Information Section S8.Source: Section S1. Materials and Methods · p. 12

Starting Materials; Analytical Techniques and Instruments

• The methods list the purchased reagents and solvents and describe elemental analysis, PXRD, NMR, TGA, and nitrogen adsorption measurements.Nitrogen adsorption isotherms were recorded at 77 K.Source: Section S1. Materials and Methods · pp. 12–13

X-Ray Diffraction Collection

• Single-crystal X-ray diffraction procedures are reported for MOF-521-H and -mF using a Rigaku instrument, and for MOF-521-oF at the Advanced Light Source.The diffraction datasets were processed with the software and correction procedures specified for each measurement.Source: Section S1. Materials and Methods · p. 13

H3BTB-oF

• Reported yields were 71% for BTB-H, 91% for the triboronic ester precursor, 52% for BTB-oF, 42% for BTB-mF, and 68% for BTB-CH3.The linker procedures use palladium-catalyzed coupling and hydrolysis steps, with conditions varying by linker.Source: H3BTB-oF · pp. 14–16

S2.2. Synthetic Procedures of Metal-Organic Frameworks

• The optimized MOF-521-H, -oF, -mF, and -CH3 procedures use individual reaction conditions, including different temperatures, durations, washing schedules, and activation conditions.All four procedures use aluminum nitrate, the corresponding H3BTB linker, formic acid, water, and DMF.Source: S2.2. Synthetic Procedures of… · p. 17
• Manual synthesis under the described conditions produced a mixture of MOF-521-H needle crystals and MOF-520 block crystals.The synthesis used 140 °C heating for 3 to 5 days.Source: S2.2. Synthetic Procedures of… · p. 17
• Elemental-analysis results are reported against calculated compositions for each of the four activated MOF-521 variants.The reported measurements include found carbon and hydrogen percentages.Source: S2.2. Synthetic Procedures of… · p. 17
• Table S1 covers screening conditions for synthesis of MOF-521-H, -oF, -mF, and -CH3; its notes define modulator abbreviations and the linker-to-metal ratio.The screening table is identified as conditions guided by the Reticular ChemNavigator through iterative learning and inference.Source: Table S1 · p. 18, S2.2. Synthetic Procedures of… · p. 19

1H NMR Spectroscopy

• BTB-X linkers were dissolved in DMSO-d6 with sonication, and their 1H NMR spectra were collected on a Bruker NEO-500.The supporting figures show spectra for H3BTB-H, -oF, -mF, and -CH3.Source: Section S3. 1H NMR Spectroscopy · p. 20, Figure S7 · p. 20, Figure S8 · p. 20
• The section reports linker and digested-MOF NMR measurements using Bruker NEO-500 and AVB-400 instruments, respectively.The linker measurements used DMSO-d6; the MOF measurements used digested activated samples.Source: Section S3. 1H NMR Spectroscopy · pp. 20, 22
• Activated MOF-521-X samples were exchanged with DMF and acetone, evacuated, then digested in 5% NaOD in D2O for 1H NMR analysis.Supporting figures show the digested MOF-521-H, -oF, -mF, and -CH3 samples.Source: Section S3. 1H NMR Spectroscopy · p. 22, Figure S11 · p. 22, Figure S12 · p. 22

Section S4. Thermogravimetric Analysis

• No significant weight loss occurred before 260 °C across all MOF-521 compounds during TGA under nitrogen flow.The temperature ramp was 5 °C per min from room temperature to 800 °C.Source: Section S4. Thermogravimetric Analysis · p. 24
• TGA traces are provided for activated MOF-521-H, MOF-521-oF, MOF-521-mF, and MOF-521-CH3 under nitrogen flow.Source: Figure S15 · p. 24, Figure S16 · p. 24, Figure S17 · p. 25

Section S5. Additional Powder X-Ray Diffraction Measurements

• The PXRD patterns of MOF-521 compounds are shown alongside a simulated pattern generated from the MOF-521-H single-crystal structure.Source: Figure S19 · p. 26
• PXRD patterns are shown for MOF-521-H soaked in pH 2, 7, and 10 solutions at room temperature.Source: Figure S20 · p. 27
• PXRD patterns compare as-synthesized MOF-521-H with MOF-521-TS, the sample calcined at 450 °C under nitrogen in the TGA furnace.Source: Figure S21 · p. 27

Section S6. Nitrogen Sorption Isotherm Analysis

• 1696 m² g-1 for MOF-521-H, 1562 m² g-1 for MOF-521-oF, 1535 m² g-1 for MOF-521-mF, and 1311 m² g-1 for MOF-521-CH3 are the reported SBET values.Source: Figure S22 · p. 28, Figure S23 · p. 28, Figure S24 · p. 29
• 10.8 Å for MOF-521-H, 10.7 Å for MOF-521-oF, 10.9 Å for MOF-521-mF, and 9.3 Å for MOF-521-CH3 are the reported pore widths.Pore-size distributions were fitted to the adsorption branches using a QSDFT model of N2 on carbon at 77 K.Source: Figure S26 · p. 30, Figure S27 · p. 30, Figure S28 · p. 31
• N2 sorption isotherms for all four variants are reported at 77 K, with filled circles for adsorption and empty circles for desorption.Source: Figure S30 · p. 32, Figure S31 · p. 32, Figure S32 · p. 33

Section S7. Single Crystal X-Ray Diffraction Analyses

• Processed crystal data were solved by SHELXT intrinsic phasing and refined by full-matrix least squares on F2 in SHELXL using Olex2.A Platon SQUEEZE SWAT correction addressed diffusing solvent in the void.Source: Section S7. Single Crystal X-Ray… · p. 34
• Crystal-data and structure-refinement tables and asymmetric-unit depictions are provided for MOF-521-H, MOF-521-oF, and MOF-521-mF.Source: Table S2 · p. 34, Figure S34 · p. 35, Table S3 · p. 36
• All three refinements report peripheral BTB-ring disorder, an unresolved water-or-DMF identity for solvent oxygen O4, and inversion twinning.Source: Section S7. Single Crystal X-Ray… · pp. 35, 37, 39

Section S8. Prompt Engineering

• The ChemScope prompt combines a role, reticular-chemistry literature, project details, and reasoning guidance to request a research-stage blueprint.Users can adjust the prompt and literature input for a specific focus, then interact with ChemScope to refine the stage content.Source: Section S8. Prompt Engineering · p. 40, Figure S37 · p. 40, Figure S38 · p. 41
• The ChemNavigator prompt uses prior project summaries, including successes and failures from earlier MOF-521-X projects, to inform later evaluations and recommendations.The first project lacks a prior summary; subsequent projects can draw on the preceding project's summary.Source: Section S8. Prompt Engineering · p. 42
• The Navigator output summary preserves project history while adding recent task outcomes and stage status, with a 30-sentence limit and explicit stage-advance rules.The stage advances only when the human explicitly says they are ready; otherwise, the iteration number increases.Source: Section S8. Prompt Engineering · pp. 42–43
• The ChemExecutor prompt combines finalized, human-refined project stages, selected-task instructions, and prior-stage summaries to generate detailed procedures.Its memory input can omit redundant iterations, using the last iteration from each stage when appropriate.Source: Section S8. Prompt Engineering · p. 44
• ChemExecutor does not require structured outputs because task procedures and reporting templates vary, while Navigator tasks can extend beyond experiments to analysis and literature review.The example task instructions specify mixed-modulator trials, hold selected reaction conditions constant, and assess products by PXRD.Source: Section S8. Prompt Engineering · p. 45, Figure S41 · p. 46
• Interactive prompt refinement cycles through Agent 1 drafting, Agent 2 testing, human evaluation, and revision until the output meets expectations.A test exposed incorrect stage-and-iteration updating, prompting more explicit instructions and examples in the revised prompt.Source: Section S8. Prompt Engineering · p. 47, Figure S44 · p. 49, Figure S46 · p. 50

Section S9. Human-AI Interaction.

• A Python script can automate prompt generation between iterations, but the authors state that using it is optional.Without it, information from a prior iteration must be extracted and incorporated into the next prompt.Source: Section S9. Human-AI Interaction · p. 51
• The authors evaluated 102 ChemNavigator task suggestions across 34 MOF-521-H iterations using three criteria scored up to three points in total.Each criterion received one point when met and zero otherwise.Source: Section S9. Human-AI Interaction · p. 51
• One criterion checked whether a suggested task coherently built on recent findings, observations, or feedback.The relevance description emphasizes consistency with, or refinement of, previously determined synthesis and characterization conditions or steps.Source: Section S9. Human-AI Interaction · p. 51

2. Potential for Progress

• A recommended task should advance the project by refining current methods or initiating a new research stage.Source: §2 · p. 51
• The authors describe progress as a form of accuracy in research progression, linking recommendations to the study's logical evolution.Source: §2 · p. 51
• Redundant suggestions that do not fit the study's progression may be less accurate guides to next steps.Source: §2 · p. 51

3. Helpfulness

• Over 80% of proposed tasks were pragmatic and geared toward research advancement; helpfulness asks whether tasks are clear, detailed, feasible, and useful for subsequent phases.The reported scores cover potential for progress and helpfulness.Source: §3 · pp. 51–52
• NMR confirmed successful BTB linker synthesis using alternative reagents, allowing the project to proceed to MOF screening and synthesis optimization.Earlier inventory checks found some required reagents unavailable, prompting validation of suitable alternatives.Source: §3 · pp. 56, 58, 59
• Formic acid was favored after acetic and trifluoroacetic acid reduced PXRD peak intensity and broadened peaks; adding water to formic acid showed promise for enhancing crystallinity.Benzoic acid produced an altered pattern, while tested formic acid:water ratios yielded similar crystallinity; a 4:1 ratio produced smaller crystals.Source: §3 · pp. 62–64
• Temperatures below 120°C did not produce MOFs, while 140°C yielded slightly more product; similar PXRD patterns across reaction times indicated structural stability and robust conditions.The results shifted subsequent work toward other synthesis factors and porosity assessment.Source: §3 · pp. 65–66
• A Type I nitrogen-sorption isotherm and successful BET analysis confirmed microporosity, while minimal adsorption–desorption differences indicated pore-structure stability.The apprentice initially needed further instruction interpreting isotherms and applying BET analysis.Source: §3 · pp. 69–72
• SCXRD findings, together with PXRD results, supported MOF-521-H as a new compound rather than known MOF-520 after its pattern failed to match literature data.The XRD mismatch prompted further structural analysis, including SCXRD, to assess structure and phase purity.Source: §3 · pp. 74–76
• TGA indicated MOF-521-H thermal stability to around 300°C, with a second mass-loss drop at 630°C.The first drop suggested decomposition onset; the second may reflect breakdown of different structural components.Source: §3 · p. 77

Helpfulness

• MOF-521-H was water-unstable, and NMR after dissolution in D2O/NaOD detected formate; CHNS results matched calculations for formate-containing MOF, implicating formate in instability.Partial solubility in D2O and inability to dissolve in DMSO-d6 complicated NMR preparation; the formate-water instability link remained a suspicion to investigate.Source: §3 · pp. 79–83
• Reproduction of MOF-521-H under optimized conditions gave a congruent PXRD pattern, supporting protocol reliability and consistent high porosity, thermal stability, and chemical stability.The matching initial and repeat characterization results supported proceeding to final validation.Source: §3 · pp. 85–86
• After an initially uncertain NMR interpretation, successful BTB-oF linker synthesis was supported by correlation between experimental and predicted 1H NMR shifts.A literature search found an exact synthesis route, which the apprentice checked for safety and feasibility before carrying out.Source: §3 · pp. 91–95
• Initial BTB-oF MOF syntheses produced white powders, and PXRD indicated that the tested BTB:Al ratios were not yet optimal.Varying formic acid:water modulator ratios improved the PXRD result, but peaks remained weak before further optimization.Source: §3 · pp. 98–100
• With formic acid:water at 4:1 and BTB:Al at 3:4, PXRD favored MOF formation; 120°C improved peaks, while higher temperatures yielded crystals and powder and extending 72 to 96 hours added little.The 4:1 modulator initially gave only weak PXRD peaks; 120°C was the apparent optimum.Source: §3 · pp. 100–103

MOF-521-oF characterization

• MOF-521-oF activation and nitrogen sorption established permanent microporosity, and successful hysteresis and porosity analyses enabled progression from stage three to stage four.Early activation produced no visible change, and an initial nitrogen-sorption dataset was incomplete before later calculations confirmed microporosity.Source: §3 · pp. 105–108
• MOF-521-oF showed reduced PXRD intensity after water exposure, while NMR and SCXRD findings implicated formate hydrolysis and possible water hydrogen bonding in its instability.Reduced PXRD intensity in acidic, neutral, and basic water conditions reaffirmed the instability; reproducibility testing later replicated the synthesis-to-characterization process.Source: §3 · pp. 112, 113, 114, 115, 117

Final summary (MOF-521-mF)

• The BTB-mF linker was synthesized by adapting a BTB-oF route and confirmed by correlation between experimental and predicted 1H NMR chemical shifts.The adapted procedure used palladium-catalyzed Suzuki-Miyaura coupling followed by saponification and acidification.Source: §3 · pp. 118, 124
• Initial MOF-521-mF synthesis conditions—BTB-mF:Al 3:4, formic acid:water 4:1, and 120°C—yielded sharp PXRD peaks; changing temperature, ratio, or solvent proportions reduced crystallinity or yield.Across tested reaction times of 60, 72, and 84 hours, single crystals formed with similar PXRD patterns; 130°C favored powders, 110°C reduced yield, and ratios 3:3 or 3:5 were less crystalline.Source: §3 · pp. 118, 126, 127, 128, 129
• MOF-521-mF was microporous after methanol activation, with a 1535 m²/g specific surface area, 0.562 cm³/g total pore volume, and 10.9 Å average pore width.The nitrogen adsorption-desorption data showed a type I isotherm without significant hysteresis.Source: §3 · p. 118
• MOF-521-mF characterization identified formula Al3(μ-OH)1(HCOO)3(BTB-mF)1, thermal stability up to 300°C, and the P-62c space group with peripheral phenyl-ring disorder.Water exposure suggested hydrolysis, while the MOF remained stable in DMF, MeOH, EtOH, and humid conditions.Source: §3 · p. 118

• Stage 2 experiments supplied substantial data on synthesis parameters despite suboptimal formic acid-to-water adjustments, and successful synthesis and screening determined optimal conditions.Source: §3 · pp. 130–131
• MOF-521-mF showed microporosity: nitrogen sorption values matched expectations, and its type I isotherm lacked hysteresis, supporting reversible adsorption and successful activation without structural loss.High-throughput screening established optimal synthesis conditions for proceeding to activation and porosity analysis.Source: §3 · pp. 131–134
• CHNS matched theoretical composition, TGA indicated stability to 300°C with subsequent stability to around 630°C, and SCXRD identified rod SBUs despite positional disorder.The SCXRD structure contained AlO6 octahedra linked through μ2-OH, with BTB-mF carboxylates and formates at the remaining corners.Source: §3 · pp. 136–138
• MOF-521-mF was stable in tested organic solvents and under humidity, but liquid water caused hydrolysis, distinguishing air stability from potential water instability.NaOD enabled dissolution in D2O for 1H NMR analysis, where a formate peak at ~8.5 ppm confirmed hydrolysis.Source: §3 · pp. 139, 141, 142
• Replicated MOF-521-mF synthesis reproduced the PXRD pattern, while CHNS, TGA, and SCXRD results from a new sample matched previous data.These results supported the reproducibility of both synthesis and characterization protocols.Source: §3 · pp. 144–145
• BTB-CH3 linker synthesis succeeded using procedural modifications and reagent substitutes, and scaled-up synthesis yielded significant product.The linker synthesis was confirmed by TLC and NMR analysis.Source: §3 · pp. 150–151
• At 130°C, MOF-521-CH3 yield increased with a similar PXRD pattern, but longer reaction time produced no further gain; 120°C trials gave little product.The initial 3:4 BTB-CH3:Al solvothermal synthesis at 120°C produced a PXRD pattern matching the predicted MOF pattern, but little product.Source: §3 · pp. 153–156

3. Helpfulness

• The optimal BTB-CH3:Al ratio was 3:4; a 4:4 ratio yielded no PXRD peaks, while higher temperatures reduced crystallinity and supported 130°C as optimal.The status evaluation suggested further tuning reaction time and solvent composition.Source: §3 · pp. 157–158
• Activation produced a pure, desolvated MOF, and sorption analysis indicated expected porosity and a stable, homogeneous microporous structure without significant hysteresis.The initial activation assessment noted that nitrogen sorption isotherms still required detailed interpretation.Source: §3 · pp. 160–162
• CHNS analysis suggested a 1:3 formate-to-BTB-CH3 ratio, and TGA showed an initial decomposition temperature of 330°C consistent with formate loss and thermal stability.Source: §3 · pp. 164–165
• SCXRD failed because the crystals were small and rod-shaped, and CD3OD insolubility prevented 1H NMR analysis; digestion in NaOD 5% in D2O revealed a formate peak at ~8.5 ppm.Source: §3 · pp. 166–168
• MOF-521-CH3 was stable in DMF, methanol, and ethanol but not water, where soaking reduced PXRD peak intensity.Source: §3 · p. 169
• The apprentice replicated linker and MOF synthesis, with 1H NMR and PXRD consistent with initial rounds, and reproduced activation and porosity results.Similar BET surface area, pore width, and pore volume supported reproducibility of the activation process.Source: §3 · pp. 171–172